This still relies on visual feedback (you need to move a cursor on a screen). It would be better if you just had to think the letter itself, not drive a cursor over it.

Not much of a difference, actually. Once you've got your brain trained to select the letters by location you can remove the eye tracking part and just use abstract thought.

"I'm thinking of a letter... On the far left of the keyboard, middle row."

I think there's not enough bandwidth for such information density. These usually work with 2D directional commands, you're driving the cursor up/down/left/right. You're not thinking of "letter E", but look at the screen and think of the direction where the cursor has to move from its current location to drift to E.

You can use auditory feedback instead of visual feedback to drive the system. Even if you are down to say 2 words per minute that's vastly better than zero.